Everything worth
asking first.

Our own coding model, trained end to end on software engineering — reading repositories, writing diffs, calling tools and reading what came back. It powers everything in Kaeonix: the chat surface, the terminal agent, and the API. The version serving today is Kaeonix Avatar 1.0, and there is no model picker anywhere — the agent and the model are tuned together, and a menu of engines is what stops that being possible.

A general model learns coding as one skill among thousands. Kaeonix Avatar was built around the engineering loop from the start, so tool-calling is first-class: it navigates a codebase, edits files and runs commands as an agent rather than describing what you should do next.

Enough for real work — many files, long diffs and accumulated tool output in one session. Cached context is charged at a reduced weight, so a long session costs considerably less than its raw token count suggests, and a session that fills the window compacts rather than stops: it is summarised and continues from the summary.

As long as the effort dial asks it to. The dial sits in the composer in chat and behind /effort in the terminal: its lowest stop skips deliberation entirely for a fast answer, and its highest asks for the model's deepest reasoning on work that earns the wait. In chat the dial belongs to the conversation, so a quick question does not re-dial the thread in your other tab.

Kaeonix is the platform: the model, the account, the billing and the surfaces you reach it through. Kaeonix Code is the terminal agent — the one that works inside your repository.

Kaeonix Chat is the browser surface for reading, planning and asking. It shares your account and balance with Kaeonix Code, so one plan covers both without separate setup, and conversations live on the account rather than in one browser — a thread started on a laptop opens on a phone with its history and its own dials.

One line, per platform:brew install feyncode/tap/kaeonixcurl -fsSL https://install.kaeonix.com | shirm https://install.kaeonix.com/install.ps1 | iexEvery release also publishes prebuilt binaries with checksums, and a Windows installer, if you would rather not pipe a script.

Yes. An API key issued by us gives Kaeonix Avatar to your editor, your pipeline, or any agent framework you already run — it speaks the OpenAI chat-completions shape, so most clients need only a base URL and a key. There is no IDE extension: Kaeonix Code runs in any terminal, including the one inside your editor.

Model work — the tokens Kaeonix Avatar reads and writes across chat and terminal sessions. Cached context is charged at a lower weight, which is why long sessions stretch further here.

Only in headroom. Every tier runs the full Kaeonix Avatar model with the same tools; what changes is how hard you can push one session and how much you can run in a month. Pro starts at 2M tokens a session window and 100M a month; Max is the largest, at 10M and 500M.

Sessions slow rather than stop. You can move up a tier or switch to API billing mid-month, with the difference prorated.

Yes — the Kaeonix Avatar API is metered rather than pooled, on the same account. At the input rate a dollar buys 1M tokens.

Yes. The terminal agent shows live context use and cost as a session runs, and the console keeps the usage record for the account afterwards.

Max covers small teams sharing agent workloads. For seats, invoicing and pooled volume, talk to us.

No. One FeynCode account covers chat, the terminal agent, the API and the console, and one balance is drawn on by all of them.

Run kaeonix login. The terminal shows a short code, the browser confirms it against the account you are already signed into, and the credential is stored in your system keyring — never in a dotfile in your repository.

Up to 5 at once. A device is a browser session or a Kaeonix Code sign-in, and both draw on that one cap — a subscription is a pool of model work, and nothing else counted how many places were drawing on it. Sign a device out from the account settings to free a slot.

No. Kaeonix runs on our own model, Kaeonix Avatar, so access comes from a plan or an API key we issue. That is what lets us tune the agent and the model together.

No. Prompts and completions are processed in memory on GPUs FeynCode controls, there is no third-party model provider behind the API, and we do not train on your prompts or outputs. The privacy policy states it in full, including what is stored and what is not.

Only what you allow. File edits, shell commands and anything network-mutating run behind a permission layer you configure per project, rules you set once are remembered, and plan mode stays read-only until you approve the change set it drafts.

Kaeonix Code executes locally against your working tree; the model runs on our infrastructure, so what it sees is what a turn sends. It reads files on demand rather than uploading the repository, and a file it never opened is never in a request.

The status page answers that first — it carries live probes of every surface and the incident history behind them. If a turn fails while the model is under load the surface says so, and the same message can be sent again.